def _preprocess(self,
Xw,
Xw_len,
Xc,
Xc_len,
scope="preprocess_layers",
reuse=False):
with tf.variable_scope(scope, reuse=reuse):
Xw_embedded, size_w = embedded(Xw,
self.embeddings_w[0],
self.embeddings_w[1],
self.config.wv_config["train_w"],
scope="embedded_w")
Xc_embedded, size_c = embedded(Xc,
self.embeddings_c[0],
self.embeddings_c[1],
self.config.wv_config["train_c"],
scope="embedded_c")
batch_size, seq_len = tf.shape(Xw)[0], tf.shape(Xw)[1]
Xc_embedded = tf.reshape(Xc_embedded,
shape=[batch_size * seq_len, -1, size_c])
Xc_len = tf.reshape(Xc_len, shape=[
batch_size * seq_len,
])
Xc_embedded, size_c = bi_gru(Xc_embedded, Xc_len,
(self.config.char_dim, ), 2,
self.initializer, 1.0, "bi_gru_c2w")
Xc_embedded = tf.reshape(Xc_embedded,
shape=[batch_size, seq_len, size_c])
X_embedded = tf.concat([Xw_embedded, Xc_embedded], axis=-1)
out_w, out_w_size = bi_gru(X_embedded, Xw_len,
(self.config.bi_dim, ), 1,
self.initializer, 1.0, "bi_gru__wc")
return out_w, out_w_size
def _encode(self, Xw, Xw_l, Xc, Xc_l, scope="encode_layers", reuse=False):
with tf.variable_scope(scope, reuse=reuse):
Xw_embedded, size_w = embedded(Xw,
self.embeddings_w[0],
self.embeddings_w[1],
self.config.wv_config["train_w"],
scope="embedded_w")
Xc_embedded, size_c = embedded(Xc,
self.embeddings_c[0],
self.embeddings_c[1],
self.config.wv_config["train_c"],
scope="embedded_c")
batch_size = tf.shape(Xw)[0]
# char
v0, v0_size = attention_han(Xc_embedded, self.config.un_dim,
self.initializer, "attention_han_c")
v1, v1_size = bi_gru(Xc_embedded, Xc_l, (self.config.bi_dim, ), 2,
self.initializer, 1.0, "bi_gru_c")
char_v = tf.reshape(tf.concat([v0, v1], axis=-1),
[batch_size, v0_size + v1_size])
# word
v0, v0_size = attention_han(Xw_embedded, self.config.un_dim,
self.initializer, "attention_han_w")
v1, v1_size = bi_gru(Xw_embedded, Xw_l, (self.config.bi_dim, ), 2,
self.initializer, 1.0, "bi_gru_w")
word_v = tf.reshape(tf.concat([v0, v1], axis=-1),
[batch_size, v0_size + v1_size])
# phrase
Xp_embedded, size_p = conv_with_max_pool(Xw_embedded, (2, 3, 4, 5),
size_w // 4, False,
tf.nn.selu,
self.initializer,
"conv_w2p")
v0, v0_size = attention_han(Xp_embedded, self.config.un_dim,
self.initializer, "attention_han_p")
v1, v1_size = bi_gru(Xp_embedded, Xw_l, (self.config.bi_dim, ), 2,
self.initializer, 1.0, "bi_gru_p")
phrase_v = tf.reshape(tf.concat([v0, v1], axis=-1),
[batch_size, v0_size + v1_size])
return char_v, word_v, phrase_v
def build_graph(self):
self.graph = tf.Graph()
with self.graph.as_default():
with tf.variable_scope("placeholders"):
self.X1w = tf.placeholder(dtype=tf.int32,
shape=[None, None],
name="sent1w_ph")
self.X2w = tf.placeholder(dtype=tf.int32,
shape=[None, None],
name="sent2w_ph")
self.X1c = tf.placeholder(dtype=tf.int32,
shape=[None, None, None],
name="sent1c_ph")
self.X2c = tf.placeholder(dtype=tf.int32,
shape=[None, None, None],
name="sent2c_ph")
self.y = tf.placeholder(dtype=tf.int32,
shape=[
None,
],
name="label_ph")
self.keep_prob = tf.placeholder_with_default(
1.0, shape=[], name="keep_prob_ph")
self.X1w_mask = tf.sign(self.X1w, name="sent1w_mask")
self.X2w_mask = tf.sign(self.X2w, name="sent2w_mask")
self.X1c_mask = tf.sign(self.X1c, name="sent1c_mask")
self.X2c_mask = tf.sign(self.X2c, name="sent2c_mask")
self.X1w_l = tf.reduce_sum(self.X1w_mask,
axis=-1,
name="sent1w_len")
self.X2w_l = tf.reduce_sum(self.X2w_mask,
axis=-1,
name="sent2w_len")
self.X1c_l = tf.reduce_sum(self.X1c_mask,
axis=-1,
name="sent1c_len")
self.X2c_l = tf.reduce_sum(self.X2c_mask,
axis=-1,
name="sent2c_len")
X1_f, X1_b = self._preprocess(self.X1w,
self.X1w_l,
self.X1c,
self.X1c_l,
scope="preprocess_layers")
X2_f, X2_b = self._preprocess(self.X2w,
self.X2w_l,
self.X2c,
self.X2c_l,
scope="preprocess_layers",
reuse=True)
with tf.variable_scope("match_layers"):
# Shapes: (batch_size, num_sentence_words, 8*multi-perspective_dims)
match_1_to_2_out, match_2_to_1_out = bilateral_matching(
X1_f, X1_b, X2_f, X2_b, self.X1w_mask, self.X2w_mask,
self.keep_prob, self.config.mp_dim)
# Aggregate the representations from the matching functions.
with tf.variable_scope("aggregate_layers"):
seq_1_fb, _ = bi_gru(match_1_to_2_out, self.X1w_l,
(self.config.bi_dim, ), 2,
self.initializer, 1.0, "bi_gru")
seq_2_fb, _ = bi_gru(match_2_to_1_out,
self.X2w_l, (self.config.bi_dim, ),
2,
self.initializer,
1.0,
"bi_gru",
reuse=True)
combined_aggregated_representation = tf.concat(
[seq_1_fb, seq_2_fb], -1)
with tf.variable_scope("fc_layers"):
h = tf.nn.dropout(combined_aggregated_representation,
keep_prob=self.keep_prob)
h = tf.layers.dense(h,
self.config.un_dim,
activation=tf.nn.selu,
kernel_initializer=self.initializer)
h = tf.nn.dropout(h, keep_prob=self.keep_prob)
pi = 0.01
self.logits = tf.layers.dense(
h,
1,
kernel_initializer=self.initializer,
bias_initializer=tf.constant_initializer(-np.log((1 - pi) /
pi)))
self.pos_prob = tf.nn.sigmoid(self.logits)
self.var_list = [v for v in tf.global_variables()]
if self.config.fine_tune:
self.var_list_trainable = [
v for v in tf.trainable_variables()
if "embedded" in v.name or "fc" in v.name
]
else:
self.var_list_trainable = [v for v in tf.trainable_variables()]
with tf.name_scope("Loss"):
self.loss_op = build_loss(labels=self.y,
logits=self.logits,
focal=self.config.focal,
alpha=self.config.alpha,
gamma=self.config.gamma)
with tf.name_scope("Optimize"):
self.adam_op = tf.train.AdamOptimizer(learning_rate=self.config.init_learning_rate). \
minimize(self.loss_op, var_list=self.var_list_trainable)
self.sgd_op = tf.train.MomentumOptimizer(learning_rate=self.config.init_learning_rate, momentum=0.9). \
minimize(self.loss_op, var_list=self.var_list_trainable)
with tf.name_scope("Prediction"):
self.predicted = tf.cast(tf.greater_equal(
self.pos_prob, self.config.threshold),
dtype=tf.int32)
with tf.name_scope("Summary"):
self.summaries = build_summaries()